| models: | |
| - model: CultriX/Qwen2.5-14B-Wernickev3 | |
| parameters: | |
| weight: 0.38 # Slight reduction to balance with FinalMerge's generalist capabilities. | |
| density: 0.65 # Retain significant parameters for stability and strong task performance. | |
| - model: CultriX/Qwen2.5-14B-FinalMerge | |
| parameters: | |
| weight: 0.32 # Slight increase to ensure its generalist capabilities are fully utilized. | |
| density: 0.60 # Balanced density for comprehensive task coverage. | |
| - model: CultriX/Qwen2.5-14B-Emergedv3 | |
| parameters: | |
| weight: 0.20 # Retains focused contribution to specific task optimizations. | |
| density: 0.55 # Moderate density ensures efficient parameter usage. | |
| - model: qingy2019/Qwen2.5-Math-14B-Instruct | |
| parameters: | |
| weight: 0.10 # Consistent with its specialist focus, balancing lower weight with higher density. | |
| density: 0.70 # High density ensures retention of advanced reasoning and MATH-related parameters. | |
| merge_method: dare_ties | |
| base_model: CultriX/SeQwence-14Bv1 | |
| parameters: | |
| normalize: true # Ensures all models are scaled to compatible parameter ranges. | |
| int8_mask: true # Optimizes memory and computational efficiency without accuracy loss. | |
| dtype: bfloat16 # Provides better memory efficiency and numerical stability. | |
| adaptive_merge_parameters: | |
| task_weights: | |
| tinyArc: 1.3 # Slight reduction to balance with generalist contributions. | |
| tinyHellaswag: 1.3 # Maintains strong performance in contextual reasoning. | |
| tinyMMLU: 1.2 # Balanced focus for domain-specific knowledge. | |
| tinyTruthfulQA: 1.2 # Adjusted to ensure fair contribution without over-prioritization. | |
| tinyTruthfulQA_mc1: 1.1 # Maintains a moderate priority to balance with other tiny benchmarks. | |
| tinyWinogrande: 1.2 # Strong contextual reasoning support from generalist models. | |
| IFEval: 1.5 # High weight for general instruction-following capabilities. | |
| BBH: 1.5 # Prioritizes complex reasoning and multi-step problem-solving tasks. | |
| MATH: 1.55 # Slight reduction to balance MATH with other advanced reasoning benchmarks. | |
| GPQA: 1.4 # Balanced to reflect contributions from both generalist and specialist models. | |
| MUSR: 1.4 # Increased slightly to strengthen multi-step reasoning. | |
| MMLU-PRO: 1.3 # Maintains general task performance across multitask domain knowledge. | |
| smoothing_factor: 0.18 # Slightly increased for smoother blending across task boundaries. | |
| gradient_clipping: 0.88 # Tightened slightly for stability, preventing parameter over-contribution. | |